In Small-Business Credit, The AI Question Isn’t Settled And The Operators Disagree

After more than a decade of putting AI to work on small-business loans, the verdict is split. New 2025 Federal Reserve data and a regulatory rewrite are forcing the industry to pick a side.

getty

Walk through a working-capital file at any non-bank small-business lender today. Almost nothing in it has been read by a person before it reaches the underwriter’s desk. A connector pulls the borrower’s books from QuickBooks. A model reconciles two years of bank statements and pings on a sudden revenue dip — the kind that, nine times out of ten, traces back to one slow-paying customer rather than a real problem. Another script runs identity checks on the personal guarantor, confirms the operating entity, and pulls a lien on the equipment offered as collateral. Two hours in, the only step left is the credit call itself.

That last step is where the argument lives. After more than a decade of putting machine learning to work on small-business credit, the people who do this for a living are not split over whether AI belongs in the building. That fight is over. They are split over the desk where the file finally lands. Who signs?

The Record Is Messier Than The Headline

Take a quick look at the last five years and the case for AI underwriting looks ugly. Kabbage, the Atlanta lender that spent a decade telling a story about machine-learning credit scoring beating banks, was sold to American Express in August 2020. The deal explicitly excluded the existing loan book. The carved-off servicer, doing business as KServicing, filed for Chapter 11 in October 2022 while the Justice Department investigated its PPP origination practices. OnDeck, the other marquee name, went to Enova International for roughly $122 million in October 2020, a long way from its 2014 IPO valuation. Upstart, briefly the standard-bearer for the whole category, lost more than 90% of its market value through 2022.

That’s the easy story. It is also wrong by half.

Upstart finished 2025 with $1 billion in revenue, up 64% year over year, and $53.6 million in net income, its strongest year on record, in a macro environment harder than its 2021 peak. The company says 91% of its 2025 loans required no human involvement. In February 2026 it named co-founder Paul Gu to replace Dave Girouard as chief executive on May 1, 2026.

The Kabbage story is more complicated than the headline reads, too. American Express took the technology, the team and the data platform; the bankruptcy was the leftover. Whether that vindicates the underwriting model or its limits depends on which side of the carve-out you count.

The honest read of the record: the cycle bit hard. Fintech lenders without much capital cushion got chewed up first. At least one of the survivors has come back to genuine profitability. “AI underwriting failed” is not what the record shows. Neither is “AI underwriting won.”

Two Camps, Same Problem

Among lenders still in the game, the question now is where to draw the line once AI is in the building.

The first camp uses AI to make the credit decision itself. Upstart’s published playbook is the cleanest version. In an SEC filing released the same week its stock cratered in August 2022, the company laid out vintage-level default data for nearly two million loans and argued that its model separated risk roughly five times more sharply than traditional credit scores. The position then and now: a properly trained model, fed the right data, beats a person staring at the same file. The 91% automation rate is what that belief looks like in production.

The second camp keeps AI everywhere except the decision. Document intake, identity and business verification, collateral revaluation, fraud screening, ongoing monitoring of the book — all machine work. The credit call itself stays with a named human risk team.

Alexander Lang, CFOCFO and co-founder of Maclear AG, a Swiss-regulated SME lender that originates loans backed by physical collateral, said in an interview that the camp distinction does not turn on capability. “The AI debate isn’t really about capability. It’s about accountability” Lang said. “A model can flag risk. A person carries it.” Per Maclear’s own platform reporting, the firm has experienced one default to date, which it says was resolved with full principal recovery; those are company-reported figures and have not been independently audited.

Working Paper 1244, by Leonardo Gambacorta and coauthors, used Italian credit-register data to test what happens when banks add AI to their credit scoring. The finding: AI investments helped banks mitigate the countercyclical rent-extraction effect of long banking relationships in normal times. During the Covid shock, AI conferred no additional credit or interest-rate protection. Relationship lending did. Read plainly, AI sharpens credit when the world looks like the data the model was trained on, and falls short of a human read when it does not.

What The 2025 Fed Survey Says

The new piece of evidence both camps are wrestling with comes from the Federal Reserve. Its 2025 Small Business Credit Survey, published in March 2026 and covering 6,525 employer firms, found that applications to online lenders rose for the fifth straight year. It also found that 60% of small businesses that borrowed from online lenders said their actual borrowing costs came in higher than they had expected. The comparable figure at small banks was 37%. At large banks it was 32%.

That 28-point gap is the open question. Faster, more automated underwriting was supposed to widen access to credit and bring price discipline to a market traditional banks were walking away from. Access is up — the applications confirm that. The price part is harder to call.

Two plausible readings. One: online lenders price riskier borrowers riskier, which is what risk-based pricing does, and the 60% figure simply reflects the population of borrowers who actually qualify. Two: speed and ease at origination obscure the all-in cost until repayments begin, and small borrowers without a finance team often do not catch the difference until after they sign.

The survey does not pick a side. It does sit awkwardly with any clean story about automation making the market work better for the people borrowing. The more useful question after sixty percent is whom AI lending is actually serving — and whom it is pricing past the point of explanation.

When AI Stops Being Expensive

The economics of running an AI credit shop have shifted hard since the Kabbage and OnDeck era. What cost millions to build in 2018 — model training pipelines, document parsers, KYB infrastructure — is now an API call at cents per query. A regional non-bank lender or a community bank can, in 2026, run the kind of intake-to-monitoring stack that used to take a forty-person engineering team to stand up.

That changes the calculus for the human-underwriting camp. Keeping AI out of the credit call is a defensible operating choice when full automation costs ten times more. It is a harder pitch when the cost gap closes.

Asked how Maclear’s approach holds up if a competitor underprices it on rate and speed, Lang argued the calculus changes once the EU AI Act takes effect. “In Europe, human accountability isn’t a philosophy — by August, it’s a regulation,” he said. “We’ve been pricing that in from day one.” The bet is that the price gap closes once compliance costs reset across the market.

Whether the bet pays off is a portfolio-level question that no single quarter or single cycle will answer. SBA 7(a) lending hit $33.4 billion in fiscal 2025, up 20.5% year over year, and SBA’s program-performance data puts the long-run charge-off rate for 7(a) in the 2% to 4% range. Hard to beat that with automation or without it. Online lenders compete in a different slice of the small-business market, but that benchmark is the one borrowers will eventually run.

The Rule Book Is Mid-Rewrite

The regulatory ground is moving fast enough that any operator who built around it five years ago is now reading new documents. In May 2025, the Consumer Financial Protection Bureau withdrew 67 guidance documents through a Federal Register notice that said the withdrawals were not final and would not be enforced while review continued. Among them was the 2022 circular that told lenders they could not justify non-compliance by pointing to model complexity.

Then, on April 17, 2026, the Federal Reserve, OCC and FDIC jointly issued SR 26-2, replacing the SR 11-7 model-risk framework that had governed bank model use for fifteen years. The new guidance is most relevant to banks with more than $30 billion in assets and places generative and agentic AI explicitly outside its scope, a footnote that points to a regulatory vacuum rather than to the absence of expected discipline.

Across the Atlantic, the EU AI Act’s high-risk obligations for credit-scoring systems take effect August 2, 2026, with documentation, bias-testing and human-oversight requirements that will reach any non-EU lender serving European borrowers.

Three things are worth watching over the next year or two. SME charge-off rates by lender type, especially how Upstart-style fully automated cohorts perform against relationship-style lenders through any real downturn. What the CFPB does after its review concludes — whether explainability and adverse-action standards come back in some form, or whether the field gets handed to state regulators. And consolidation in the AI-tooling layer, which will tell whether the price collapse keeps going or whether two or three vendors take the market and lock it.

The question is still open. Whoever ends up writing the loans the banks will not will run a coalition of person and model, with the seam between them drawn in different places at different shops. Which seam holds up will be answered by the next real downturn, not by this article.